Fig. 6: Segmentation of marine debris with multi-spectral medium-resolution imagery.
From: Meta-learning to address diverse Earth observation problems across resolutions
Marine debris on a 12-band, medium-resolution (10 m) Sentinel-2 scene near Accra, Ghana (October 31, 2018) have been visually detected and annotated by Mifdal et al.6 in (a). This task-model is modified for segmentation and initialized with the weights of the meta-model. Five annotated images of (a) serve as a training dataset to fine-tune the task-model. b A test image and the corresponding prediction in (c) by the task-model to a map showing the probability of marine debris for each pixel. The same meta-model, as in the previous use-cases, achieves these predictions, highlighting the versatility of meta-learning for various Earth observation problems with few training samples.
